Sharath Reddy

Owned data quality and observability for the cQES 1.1 telemetry ecosystem, validating 200M+ daily records across 84K U.S. census tracts in GCP BigQuery and Cloud Composer to protect downstream ML feature pipelines. Optimized distributed SQL validation workloads, reducing scanned data from 604 GB to 58.6 GB by removing nested cross-joins and enforcing partition pruning via single-scan aggregations. Operationalized automated data-quality checks in DataBuck with 7-day rolling baselines and time-series anomaly detection to identify training-data drift early. Used Claude Code as an AI coding agent for a 12-KPI geospatial truncation-drift investigation, isolating a 63.3% boundary leakage rate caused by multi-decimal coordinate truncation.

Prototyped agentic AI workflows for Verizon using LLM-based planning with LangChain and enterprise data integration. Improved NPP planning by integrating knowledge graphs (GraphDB) via SPARQL with Verizon’s Gen AI service (VEGAS). Built VTrace, an internal AI-driven system for 5G network assurance and root cause analysis that automates analysis of .pcap files. Architected an MCP interface to integrate Elasticsearch as a tool, enabling agents to query Kibana logs alongside .pcap data and improve diagnostic context by 35%. Built a hybrid sparse+dense RAG pipeline over 3GPP telecom standards with hierarchical chunk-aware preprocessing and hybrid retrieval for high-precision grounding; evaluated RAG quality using RAGAS and improved faithfulness scores to 0.90+. Integrated Milvus, MongoDB, Watsonx, and LiteLLM into a FastAPI + React UI and tuned retrieval/generation across GPT-OSS-128B and Granite models.

Engineered end-to-end multimodal document-processing pipelines using LayoutLMv3 to extract text and spatial features from unstructured financial statements, deploying inference workflows through Azure ML for two banking clients. Fine-tuned FinBERT and RoBERTa and implemented cosine similarity-based matching to reduce manual review by 40%. Designed a hybrid intelligence framework combining regex heuristics with XGBoost, improving compliance audit precision by 18% and decreasing false-positive flags by 22% across Azure Synapse Analytics workloads. Optimized deep learning models with TensorFlow/Keras and Bayesian hyperparameter optimization, achieving a 0.94 F1 score across heterogeneous audit datasets and improving generalization.